// WRITING · 2026.02.10
What i'm currently breaking and rebuilding
notes from the workbench — pydantic ai pipelines, agent loops, and other things held together with tape.
A running list from the workbench. None of this is finished. That is the point — the interesting work always looks like something half-taken-apart on the bench.
On the bench right now
- Pydantic AI pipelines — pushing more of the agent’s contract into typed schemas so failures surface at the boundary instead of three tool-calls deep.
- Agent loops — replacing my hand-rolled control flow with something I can actually reason about when it loops forever at 2am.
- Eval harnesses — the small honest kind (see the last post), wired to run on every change so a “harmless” prompt tweak can’t quietly regress.
Things I broke this month
- A retrieval cache that was serving yesterday’s embeddings with total confidence.
- A memory writer that summarized so aggressively it forgot the user’s name.
- A market-data ingest that fell behind by exactly one lunch break.
Each of these taught me more than the version that worked. I break things on purpose now, on a schedule I control — because the alternative is breaking them by accident, on a schedule the user controls.